• Transportation presents a major challenge to curbing climate change.

• Better informed policymaking requires up-to-date empirical data with good quality, low cost, and easy access.

• Emerging data sources enable deep and new insights from large-scale collection of human movement and transport systems.

• Why Twitter?

  • Easy access, low cost, large spatial and population coverage.

• Limitations of geotagged tweets

  • Biased population: young, highly-educated, urban residents.

  • Sparse sampling of the actual mobility.

  • Behaviour bias of reporting geolocations.

• Local vs. Global traveller visits

  • Local: nearby locations.

  • Global: more distant locations.

• Returner vs. Explorer explores around

  • Returner: one centralised location.

  • Explorer: decentralised locations that are distant from each other.

spatial scale sampling method sample size

• Twitter data are more suitable for city level than national level.

• The main obstacle of using Twitter data at a large spatial scale is the sparsity.

Fig. 9 National level (left) vs. city level (right).Paper II

spatial scale sampling method sample size

• User-based data collection works better than area-based data collection:

  A much larger number of geotagged tweets, a more complete picture of travel demand.

• A density-based approach is proposed to increase sample size:

  

  Fig. 10 Trip-based approach (left)Lee et al., 2019 vs. density-based approach (right).Paper II

A density-based approach A mobility model

• An individual-based mobility modelPaper III fills the gaps in sparse mobility data, particularly geolocations of social media data.

• The model is designed to correct behaviour bias and sparsity issue.

A density-based approach A mobility model

• The model-synthesised results have good agreements with the other data sources.

• Characterising trip distance distributions (domestic) of global regions:

Fig. 11 Distributions of synthesised domestic trips.Paper III

Data fusion framework for travel time calculationPaper IV:

Distributions of geotagged tweets represents the dynamic attractiveness of locations in cities.

Fig. 12 Geotagged tweets as destinations by hour of day (Stockholm region).

Spatiotemporal dynamics of travel time ratio (R)Paper IV:

Fig. 13 Travel time ratio by hour of day (Sydney).

• Travel time by PT is around twice as high as by car.

• PT can compete with car use during peak rush hours in Stockholm and Amsterdam.

Fig. 14 Travel time ratio over 24 hours.

• How large is the share of ride-sourcing trips that can be substituted by taking public transit, if you are willing to walk up to 800 m to access and leave the transit station during daytime?

• Ride-sourcing trips: transit-competing vs. non-transit-competing.

  • What trip attributes and built environment are linked to the competition?

  • What are the implications for policymaking?

For travel time calculation For open trip data analysis (ride-sourcing)

• Collected from a large area and population but at a cost of rich detail.

• Raw data: trip ID, pick-up and drop-off locations, pick-up and drop-off times, and cost.

A glass-box model enhanced by machine learning techniques: additive impact of factors and factor interactions.

• The transit-competing trips account for 48.2%.

• Competition is more likely to happen when the travel time by ride-sourcing < 15 min.

• Requiring multiple transfers is associated with the competition between.

Land-use clusters of the study area.

Fig. 18 Impact of land-use cluster and its interaction with transfer.Paper V

• Low density/diversity of land use -> a lower probability of competition.

• Multiple transfers + middle density/diversity of land use -> a higher probability of competition.